Mircea Neagovici — Robotic Process Automation (RPA) and ML

Fine-Tuning

Fine-tuning machine learning models to meet specific business needs is a complex process. Mircea Neagovici explains that while some models, like those for document understanding, are retrainable, others require customer feedback to improve. He highlights the preference for PyTorch over TensorFlow due to its debuggability and performance benefits, which aids in developing models faster 1. Neagovici notes, "We don't allow the computer vision model to be trained by customers or the OCR, although for the OCR, we have to get the feedback from the customers and improve" 2. This approach ensures models are fine-tuned effectively, balancing customer needs with technical capabilities.

   

Data Drift

Managing data drift is a significant challenge in machine learning. Neagovici discusses the importance of monitoring data quality before training to prevent costly errors later 3. He emphasizes the need for active learning and pretraining to enhance model performance. "We have to do more active learning. We do mostly supervised learning," he states, highlighting the ongoing efforts to improve data handling practices 4. This proactive approach aims to mitigate data drift and ensure robust model performance.

   

Performance Balance

Balancing model performance with business needs involves strategic trade-offs. Neagovici explains that sometimes simpler models are preferable, especially when dealing with similar documents 5. He mentions the importance of confidence scores, which help determine when human intervention is necessary. "Everybody will take a model that's like three points worse in terms of quality if the confidence scores are perfect," he notes, underscoring the value of reliability over sheer accuracy 6. This balance ensures that models are not only effective but also practical for real-world applications.